Reconstructing Communication Networks with Large Language Models: Receiver Inference in Relational Event History Data

This article introduces an open-source project maintained by myrthe12 (GitHub link: https://github.com/myrthe12/Receiver-Inference-in-Relational-Event-History-Data). The project uses large language models (LLMs) to infer message receivers from relational event history (REH) data, reconstructing complete communication networks for multi-participant dialogue scenarios (e.g., parliamentary debates) and providing a new technical tool for social network analysis and political science research.

Relational Event History (REH) data is a core type in social network analysis, recording elements such as timestamps, senders, action types, and content, but it often lacks receiver information. Examples include: parliamentary debate records that do not specify the target of a legislator's speech, social media tweets without clear receivers, and mailing list messages addressed to multiple receivers. This absence limits the depth of network analysis—traditional social network analysis relies on complete 'who-to-whom' relationship data, and missing receivers obscure the network structure.

The project proposes an LLM-based receiver inference framework with core components including:

1. Prompt engineering module: Provides baseline, zero-shot, few-shot, and other prompt strategies;
2. Inference module: Interacts with LLM APIs, takes context and candidate receiver lists, and returns inference results and confidence scores;
3. Evaluation module: Supports turn-level (single message prediction accuracy) and network-level (reconstructed network structure quality) evaluation;
4. Statistics and analysis module: Analyzes model confidence, uncertainty, and relational event patterns. In terms of technical details, the project mitigates long-context challenges through selective context truncation and key information extraction, handles multi-receiver issues, and quantifies model confidence.

The project is corely applied to the analysis of Dutch parliamentary debate records, enabling:

1. Reconstructing debate networks: Identifying who speaks to whom to form directed communication networks;
2. Analyzing political alliances: Identifying factions and alliance relationships from network structures;
3. Tracking issue evolution: Observing the spread and evolution of issues among legislators;
4. Quantifying influence: Evaluating legislators' influence based on network centrality.

Advantages of LLM applications in social science research:

1. Reducing manual annotation costs: Automating receiver inference lowers research costs;
2. Capturing subtle semantic cues: Identifying implicit directional signals such as pronoun usage and tone changes;
3. Transferability: Applicable to multiple scenarios like online forums, mailing lists, and social media threads;
4. Interpretability: LLMs can provide natural language explanations to help understand the inference process.

Limitations:

• Dependence on LLM bias: Biases in training data may lead to inaccurate inferences;
• Context length constraints: Long dialogue histories cannot effectively utilize all information;
• Computational cost: Inference on large datasets incurs high API costs. Future Directions:
• Multimodal extension: Combining signals like text and voice;
• Incremental learning: Quickly adapting to new dialogue domains;
• Causal inference: Inferring the impact of messages on receivers;
• Real-time applications: Applying to online meetings or live debates.

This project demonstrates the value of LLMs as tools for social science research. Automated receiver inference lowers the barrier to social network analysis, providing a usable toolkit for researchers in fields like political science and sociology. As LLM technology advances, such methods will continue to improve in accuracy and applicability, opening up new possibilities for understanding human social interactions.